Discovering frequently recurring movement sequences in team-sport athlete spatiotemporal data

ABSTRACT

Athlete external load is typically analysed from predetermined movement thresholds. The combination of movement sequences and differences in these movements between playing positions is also currently unknown. This study developed a method to discover the frequently recurring movement sequences across playing position during matches. The external load of 12 international female netball athletes was collected by a local positioning system during four national-level matches. Velocity, acceleration and angular velocity were calculated from positional (X, Y) data, clustered via one-dimensional k-means and assigned a unique alphabetic label. Combinations of velocity, acceleration and angular velocity movement were compared using the Levenshtein distance and similarities computed by the longest common substring problem. The contribution of each movement sequence, according to playing position and relative to the wider data set, was then calculated via the Minkowski distance. A total of 10 frequently recurring combinations of movement were discovered, regardless of playing position. Only the wing attack, goal attack and goal defence playing positions are closely related. We developed a technique to discover the movement sequences, according to playing position, performed by elite netballers. This methodology can be extended to discover the frequently recurring movements within other team sports and across levels of competition.

KEYWORDS:

• Netball
• WASP
• data mining
• activity profile

Introduction

Profiling the external load of team-sport athletes during matches is useful information for training design and load management (Carling, Bloomfield, Nelsen, & Reilly, 2008). External load is captured by tracking systems, including global positioning systems (GPSs), that estimate an athlete’s position with respect to the coordinates of a playing area and allow for the calculation of displacement over a specified time epoch (Aughey, 2011). Once the trajectory data of an athlete’s position is captured, the resulting velocities and accelerations can be calculated. The match activity profile of field-based team-sport athletes has been documented (Aughey, 2011); however, limited research exists on the external load of athletes participating in court-based sports. This is likely due to elite-level matches being played indoors, where GPS is inoperable. The recent development of radio frequency (RF)-based athlete tracking systems (Sathyan, Shuttleworth, Hedley, & Davids, 2012) may allow for external load to be captured during elite court-based team sports.

Athlete external load is typically presented according to the distribution of time spent or distance covered in dissimilar velocity and acceleration bands (Jennings, Cormack, Coutts, & Aughey, 2012; Varley, Gabbett, & Aughey, 2013). These predetermined thresholds are established according to the manufacturer of the tracking technology, a body of research (Varley et al., 2013) or as a function of a physiological capacity (Lovell & Abt, 2012). In female team-sport athletes, up to a 30% difference in match high-speed running was recorded between industry- (5 m · s⁻¹) and physiological capacity-based velocity thresholds (Clarke, Anson, & Pyne, 2014). Although expressing external load relative to a physiological threshold can identify athletes who are working at or near capacity, it is currently unclear which physiological tests are best representative or related to the intermittent nature of team sport (Carling, 2013). There is also limited research on the relationship between such physiological tests and the movement of athletes during court-based team sports, including netball.

A popular sport in Commonwealth countries, netball matches are contested over 15-min quarters between two teams of seven players. Each of the seven players has a unique positional role that restricts their movement to specific regions of the court. Research on the match movement output of netballers, according to playing position, is largely confined to estimates from video analysis (Fox, Spittle, Otago, & Saunders, 2013) or examined in a sub-elite cohort (Cormack, Smith, Mooney, Young, & O’Brien, 2014). Recently, the load of elite netball athletes was examined during training and matches according to playing position (Young, Gastin, Sanders, Mackey, & Dwyer, 2016). Although athlete displacement and velocity was not measured, the accelerometer-derived player load from each playing position was clustered into groups using a data mining technique (Young et al., 2016).

Data mining is a branch of computer science that sources a logical or mathematical description of patterns and regularities in a data set (Fayyad, Piatetsky-Shapiro, & Smyth, 1996). Data mining methods can extract previously unknown information from raw data and have useful application in elite sport, including the modelling and extraction of athlete performance patterns (Ofoghi, Zeleznikow, MacMahon, & Raab, 2013). Clustering is a data mining method that detects and organises data into a number of groups. The data within each of these groups or cluster are similar to one another, based on some criteria, and dissimilar to data within other groups (Ofoghi et al., 2013). In sport, clustering has been utilised to assess the tactical patterns of play during elite volleyball matches (Jäger & Schöllhorn, 2007) and classify movement patterns performed during different basketball shots (Lamb, Bartlett, & Robins, 2010). Deemed an unsupervised data mining technique, clustering does not require prior knowledge or impose boundaries on data. Data mining has also been used to extract movement activity data, including the position, velocity and acceleration of different body parts, from wearable sensors (Ghasemzadeh, Loseu, & Jafari, 2010). Each movement variable was represented by a sequence of characters, with a metric used to find the difference between two character strings (Ghasemzadeh et al., 2010).

Together, the data mining techniques of clustering and string matching present an opportunity to ascertain the movements performed by team-sport athletes, without the requirement of arbitrary or physiologically defined thresholds. Using data mining techniques, the sequences of velocity, acceleration and angular velocity performed by a junior-elite female netball athlete were examined (Sweeting, Cormack, Morgan, & Aughey, 2014). Spatiotemporal data, or the real-time position of an athlete relative to a playing area, was collected by a local positioning system. Four velocity, three acceleration and four angular velocity clusters were obtained by k-means clustering, an unsupervised data mining approach that assigns data points to a cluster based on the closest centroid (Sweeting et al., 2014). Each combination of continuous velocity, acceleration and angular velocity movement was assigned a character, with athlete movement represented by strings of characters or sequences (Sweeting et al., 2014). Eighteen movement sequences were obtained, and running in a straight or 45° direction with neutral acceleration was a common feature (Sweeting et al., 2014). However, spatiotemporal data was only collected during a quarter of netball and there was no investigation into the differences or similarities in movement sequences performed by the remaining six netball playing positions. Further, only a junior-elite athlete participated and not elite, international-level athletes. Therefore, the aim of this study was to further develop this methodology to uncover the movement sequences performed by court-based team-sport athletes, according to playing position and independent of industry-based, commercially developed or physiologically defined thresholds.

Methods

Participants

The activity profile of 12 elite, international-level female netball athletes (age 24.8 ± 2.7 years; height 179.5 ± 6.9 cm, mean ± standard deviation [SD], at commencement of study) was collected during four competitive national-level matches. The elite-level athletes who participated in this study represent their country in a limited number of international netball matches that are held annually. Therefore, only a small number of matches were sampled.

The number of individual athletes sampled per netball playing position was five for the centre (C), three for wing defence (WD), five for wing attack (WA), three for goal attack (GA), two for goal defence (GD), three for goal shooter (GS) and two for goal keeper (GK). All participants provided written informed consent. The study was approved by the University Human Research Ethics Committee (HRE14-068) and conformed to the Declaration of Helsinki.

Study protocol

Spatiotemporal data was collected via a RF tracking system, specifically, the Wireless Ad Hoc System for Positioning (WASP). Indoors, WASP has a relative positional accuracy of 28 cm and a mean distance error of 2.7% (Sathyan et al., 2012). When compared to Vicon, the WASP coefficient of variation (CV) is <6% for measuring the distance covered during five short (<15 m) non-linear courses performed indoors by elite netball athletes (Sweeting, Phillips, Morgan, Cormack, & Aughey, 2016). Over these five separate non-linear courses, the CV for WASP-derived mean velocity is <8%. For mean angular velocity, the CV is <3% and for acceleration, the CV ranges from 2.3% to 18.5% (Sweeting et al., 2016).

Each participant wore a WASP mobile node, measuring 90 × 50 × 25 mm, positioned between the shoulder blades. The range between each mobile and the 12 anchor nodes surrounding the netball court was computed at an update rate of 10 Hz and calculated into a 100-Hz file via customised software (WheresBruce, Australian Institute of Sport, Canberra, Australia). This process involves resampling the RF signal via a Kalman filter as described in Sathyan et al. (2012) and is very similar process to the RF tracking system used by Stevens et al. (2014). Further details of the Kalman filter and positioning algorithm used can be found elsewhere (Hedley, Sathyan, & Mackintosh, 2011). Each athlete’s positional data (X and Y coordinates) were exported into the R statistical software (R: A language and environment for statistical computing, Vienna, Austria) for further analysis.

Movement sequencing technique

Velocity was calculated from each athlete’s positional data and acceleration derived from velocity (Sweeting et al., 2014). Angular displacement was calculated via the dot product of consecutive movement vectors. Angular velocity, the rate of change of angular displacement, was obtained from angular displacement (Sweeting et al., 2014). Individual velocity, acceleration and angular velocity movements were clustered using a one-dimensional k-means clustering algorithm (Wang & Song, 2011) seeded with 4, 3 and 4 clusters, respectively. The cluster analysis simply assigns a data point to the nearest centroid and does not assert statistical difference; therefore, no statistical analysis was performed on these clusters. A qualitative label was assigned to each cluster, which may not align precisely with the mean values but are intended to represent arbitrary descriptors rather than specific quantities. From these clusters, each unique combination of velocity, acceleration and angular velocity movement, termed movement subunits, was assigned an identification code consisting of an upper- or lower-case alphabetic letter. In short, athlete movement during a match was represented by continuous movement subunits.

To isolate discrete athlete movement sequences, any period during which the athlete moved at a rate lower than 0.5 m · s⁻¹ was judged to be moments of inactivity and thus delineated continuous movement subunits to form sequences. The similarity between each movement sequence was quantified using the Levenshtein distance implementation in the R stringdist package (van der Loo, 2014). The Levenshtein distance represents the minimum number of movement subunits required to change, including insertions, deletions or substitutions, one movement sequence into another. Similar movement sequences were then grouped into 25 clusters using a hierarchical cluster analysis (Ward Jr, 1963).

The longest common subsequence (LCS) algorithm, using the R qualV package (Jachner, Van den Boogaart, & Petzoldt, 2007), was used to discover the most common athlete movement sequence within each of the 25 clusters. A key feature of the LCS algorithm is to discover all of the common elements or movement subunits, within movement sequences whilst retaining the sequential order. Therefore, the LCS or frequently recurring patterns of athlete movement performed across matches could be located.

To uncover the frequently recurring movement sequences performed by individual playing position, a frequency distribution was conducted using two methods. The relative frequency of individual movement subunits was compiled for each playing position. The relative frequency of the LCS-derived movement sequences for each playing position was also calculated. These distributions can be considered a movement signature for each netball playing position. Further, the Minkowski distance implemented in the R HistogramTools package (Stokely, 2015) was used to quantify the distance between playing positions using the LCS results. Briefly, the Minkowski distance was calculated by obtaining the relative percentage contribution of each LCS movement to the wider data set, for each playing position. A matrix was then constructed, with the similarity between each playing position calculated via the Minkowski distance. A network graph was used to display this similarity between playing positions.

Results

The centroids of the four velocity clusters were 0.6, 1.4, 2.5 and 3.9 m · s⁻¹, respectively. The centroids of the four angular velocity clusters were 13.5, 49.9, 98.9 and 153.6 deg · s⁻¹, respectively. Centroids of the three acceleration clusters were −6.8, 0.0 and 6.9 m · s⁻², respectively. The within-cluster variation, as the sum of Euclidean distance between the data points and each centroid, was 90.2% for velocity, 71.9% for acceleration and 94.7% for angular clusters. The distribution of data points within each velocity, angular velocity and acceleration cluster and the relative frequency of these clusters to the total data set is demonstrated in Figure 1(a–c), respectively. Qualitative labels were assigned to each cluster, which may not align precisely with the mean values but are intended to represent arbitrary descriptors rather than specific quantities.

Figure 1. The relative frequency of clustered observations for (a) velocity, (b) angular velocity and (c) acceleration movement features.

The most prevalent movement features of netball match activity were walking with straight movement and neutral acceleration. Neutral acceleration refers to acceleration data assigned to the cluster with a centroid of 0.0 m · s⁻², which is the mean of all the data points within this cluster. These data points are not necessarily accelerations of 0.0 m · s⁻². Each movement subunit, the qualitative descriptor comprising the relevant combination of velocity, acceleration and angular velocity and their relative frequency to the wider data set are presented in Table 1.

Table 1. Movement subunits, their percentage contribution to the wider data set and qualitative descriptor (this wider data set is inclusive of all players, matches and positions).

Movement subunit	Percentage contribution	Qualitative descriptor
a	6.1	Walk Neutral Turn 90
A	1.6	Walk Accelerate Turn 180
b	1.4	Walk Decelerate Turn 90
B	0.2	Run Neutral Turn 180
c	1.3	Walk Accelerate Turn 90
C	0.3	Run Decelerate Turn 180
d	0.7	Run Neutral Turn 90
D	0.3	Run Accelerate Turn 180
e	0.4	Run Decelerate Turn 90
E	1.7	Jog Neutral Turn 180
f	0.5	Run Accelerate Turn 90
F	1.0	Jog Decelerate Turn 180
g	2.5	Jog Neutral Turn 90
G	1.0	Jog Accelerate Turn 180
h	1.0	Jog Decelerate Turn 90
H	0.0	Sprint Neutral Turn 180
i	1.1	Jog Accelerate Turn 90
I	0.0	Sprint Decelerate Turn 180
j	0.1	Sprint Neutral Turn 90
J	0.0	Sprint Accelerate Turn 180
k	0.1	Sprint Decelerate Turn 90
K	8.7	Walk Neutral Straight
l	0.1	Sprint Accelerate Turn 90
L	1.5	Walk Decelerate Straight
m	6.8	Walk Neutral Turn 45
M	1.3	Walk Accelerate Straight
n	1.3	Walk Decelerate Turn 45
N	7.3	Run Neutral Straight
o	1.2	Walk Accelerate Turn 45
O	2.4	Run Decelerate Straight
p	2.4	Run Neutral Turn 45
P	2.5	Run Accelerate Straight
q	1.0	Run Decelerate Turn 45
Q	12.1	Jog Neutral Straight
r	1.1	Run Accelerate Turn 45
R	2.7	Jog Decelerate Straight
s	5.1	Jog Neutral Turn 45
S	2.6	Jog Accelerate Straight
t	1.6	Jog Decelerate Turn 45
T	3.1	Sprint Neutral Straight
u	1.6	Jog Accelerate Turn 45
U	1.3	Sprint Decelerate Straight
v	0.7	Sprint Neutral Turn 45
V	1.4	Sprint Accelerate Straight
w	0.3	Sprint Decelerate Turn 45
x	0.4	Sprint Accelerate Turn 45
y	6.7	Walk Neutral Turn 180
z	1.7	Walk Decelerate Turn 180

The 10 most frequently recurring movement sequences and the relative contribution to the wider data set, according to playing position, are presented in Figure 2. These substrings are the longest string within each of the 25 clusters. For example, in 1 of the 25 clusters, “KK” was the LCS within that cluster. In a separate cluster, “NNNNN” was the LCS within that cluster. The difference between “Q” and “QQ”, for example, is that they are one movement subunit apart and were located separately within the 25 clusters. A matrix of the frequently recurring movements across individual playing positions is displayed in Table 2. The relative proximity of playing position is visualised in Figure 3. The GD, GA and WA are the most closely related netball playing positions. The largest Minkowski distance (19.64) was between the GS and GD. The Minkowski distance between the GS and C was 19.20.

Table 2. The Minkowski distances for movement sequence distributions between netball playing positions including centre (C), goal attack (GA), goal defence (GD), goal keeper (GK), goal shooter (GS), wing attack (WA) and wing defence (WD).

	C	GA	GD	GK	GS	WA	WD
C	0.0	7.24	11.37	13.87	19.20	8.42	18.55
GA		0.0	4.55	10.19	17.23	2.58	11.60
GD			0.0	12.18	19.64	5.70	9.37
GK				0.0	7.52	7.71	10.52
GS					0.0	14.78	16.56
WA						0.0	10.30
WD							0.0

Figure 2. The relative contribution of frequently recurring LCS sequences by netball playing position including centre (C), goal attack (GA), goal defence (GD), goal keeper (GK), goal shooter (GS), wing attack (WA) and wing defence (WD). The number of iterations of each movement subunit are represented by b^N, for example K² refers to KK and N⁵ refers to NNNNN.

Figure 3. A network analysis of movement sequence similarity between playing positions.

Discussion

The purpose of this article was to describe a new analysis technique of team-sport athlete movement data. The combinations of movement performed by team-sport athletes, collected via RF athlete tracking data, were quantified by a technique that assessed for similarity or dissimilarity between playing positions. Ten frequently recurring movement sequences, across all netball playing positions and matches, were discovered (Figure 2). Only the WA, GA and GD playing positions are closely related (Figure 3). Traditional analysis of team-sport athlete external load quantifies movement as a function of commercially developed or industry-used thresholds, binning velocities or accelerations into descriptive categories. This approach does not account for variations in gender, age, position or sport. To address this problem, the present technique examines the combinations of movement performed that can then be used to underpin traditional analysis of external load, irrespective of velocity or acceleration thresholds.

Four velocity, three acceleration and four angular velocity clusters were identified via one-dimensional k-means clustering. The centroids of the velocity clusters, notionally referred to as “walking”, “jogging”, “running” and “sprinting”, are lower than the thresholds used to report the external load of female team-sport athletes during matches. For example, high-speed running performed during field hockey matches was considered as any movement over 4.19 m . s⁻¹ (Macutkiewicz & Sunderland, 2011). The amount of high-speed running completed by female rugby athletes was also underestimated when analysed according to an industry-based threshold (5 m · s⁻¹) compared with a physiologically determined (3.5 m · s⁻¹) threshold (Clarke et al., 2014). In the present study, the centroid of the “sprinting” cluster was 3.9 m · s⁻¹. However, there is no intended comparison between the movement clusters described here and industry standard thresholds. In practice, it is still difficult to compare movement thresholds derived from methods in the current study with other threshold benchmarks, because the number of clusters derived from the raw data are arbitrary. Reducing the number of initial velocity clusters would likely result in different cluster means. By establishing a standardised approach to the initial clustering stage, it would be possible to make internally consistent estimates of time in self-calibrating movement intensity zones. Therefore, rather than prescribing workload bins, the thresholds are learned directly from the data. It is important to note that these self-calibrating movement intensity zones are based on external load measures only and do not provide information relative to an athlete’s internal physiological capacity. The proposed methodology is therefore appropriate for a court-based team sport, such as netball, due to the unique court space restrictions, differing roles and anthropometric characteristics of netball playing positions (Steele & Chad, 1991). The methodology presented could also be applied to field-based team sports including soccer, rugby and Australian football although future application, including generalisation to all elite female netball athletes, would require testing on a larger data set.

The network analysis highlights that the WA, GA and GD positions are the most closely related netball positions. The GS role is characterised by movement combinations that are highly dissimilar to all the other positions. However, the most dissimilar pairwise comparison is between the GS and GD positions, an interesting finding given both playing positions are goal-based roles. In a recent study on elite netball athlete load, the GD position was grouped with the GS and GK based on the proportion of match time spent performing low-intensity activity (Young et al., 2016). The athletes in the present study comprise the national representative team, which may indicate the unique playing style of the GD position in this cohort compared with those from a lower-level competition.

Differences in the accelerometer load of netball playing positions have been investigated across state representative and recreational playing levels (Cormack et al., 2014). Higher standard (state league level) athletes performed a greater proportion of accelerometer load in the vertical plane compared to their lower (recreation)-level counterparts (Cormack et al., 2014). When comparing positions, only centre court athletes had a greater load than shooters and there were no clear differences between centres and defenders nor between shooters and defenders (Cormack et al., 2014). In the present study, the seven netball playing positions were studied at an individual level via the combination of velocity, acceleration and angular velocity movements performed by elite-level athletes. It is difficult to compare these studies in netball, given the discrepancies in methodology and tracking systems utilised. Rather than simply reporting time spent in different intensity zones, the present study presents the relative frequency of recurring movement sequences as a characteristic signature that differentiates athlete movement by playing role. Potential applications of this methodology may include a more granular evaluation of the movement output for each role than can be learned by using speed thresholds alone. This approach may be used to evaluate the developmental progress of young athletes as they build the physical attributes required to compete at more senior levels. The presented methodology may also reduce the need for the traditional collection, analysis and reporting of team-sport athlete external load in arbitrary or ill-defined movement categories. Rather than structuring training programmes around time or distance spent in these categories, practitioners and scientists can focus on training the specific movement patterns frequently performed by athletes in each playing position.

Research conducted in elite female netball has been limited to only three matches (Davidson & Trewartha, 2008; Fox et al., 2013). Future work should incorporate more matches to calculate test–retest reliability and examine the repeatability or match-to-match variability of role proximity results in netball. Analysis on the change of movement sequences performed over the duration of a match may ascertain if movements are a function of game style or an individual athlete playing within that position. Although netball was the sport analysed in this article, the methodology presented can be extended to discover the frequently recurring movements within other team sports. Utilising the Minkowski distance metric, the discovered movement features and distributions by playing role may uncover new relationships between the different playing positions in team sports. The proposed methodology may assist coaches with tactical planning, through understanding the movement sequences performed by team-sport athletes during specific match activities, for example, the movements performed in the lead-up to a shot for goal. Team-sport athletes could use also information derived from the presented methodology for performance analysis purposes, including quantitative spatiotemporal data on their angle of attack during set plays as opposed to traditional inferences from video analysis.

Conclusion

The sequences of velocity, acceleration and angular velocity movement performed by team-sport athletes during matches were discovered by a novel data mining technique. The combinations of these movement sequences were utilised to measure the strength of relationship between the netball playing positions. Using a one-dimensional k-means clustering algorithm, four velocity, three acceleration and four angular velocity clusters were obtained from netball athlete positional data. A total of 10 frequently recurring combinations of movement were discovered. To examine the relationship between netball playing positions, the percentage contribution of each frequently recurring movement pattern within the wider data set was calculated. Based on the combination of velocity, acceleration and angular velocity movements performed, it was discovered that only the WA, GA and GD playing positions are closely related. The GD and GS are the least similar netball playing positions. Future research should examine if these differences are a function of a global position or instead, each individual athlete who plays within that position. Analysis should also be extended to analyse the relationship between each playing position across playing standards, for example, a junior-elite-level GA compared to an elite GA.

Acknowledgement

The authors wish to thank the athletes and coaching staff of the Australian netball team for their participation plus support of this work.

Disclosure statement

No potential conflict of interest was reported by the authors.